A Spatial-Temporal Predictive Transformer Network for Level-3 Autonomous Vehicle Decision-Making.

Area of Science:

• Autonomous Driving Systems
• Decision-Making Algorithms
• Human-Machine Interaction

Background:

• Existing research on Level-3 autonomous vehicles (L3-AVs) lacks in-depth analysis of takeover time (TOT) mechanisms.
• Current L3-AV studies overlook spatial-temporal variations in features crucial for TOT prediction.
• TOT is not adequately considered in downstream trajectory planning for L3-AVs.

Purpose of the Study:

• To investigate the impact of takeover time (TOT) on decision-making for Level-3 autonomous vehicles (L3-AVs).
• To develop advanced models for accurate TOT prediction and trajectory forecasting.
• To enhance the safety and comfort of L3-AV decision-making processes.

Main Methods:

• Proposed an exponential smoothing transformers (ETSformer) model for TOT prediction.
• Employed a spatial-temporal predictive transformer (ST-Preformer) for surrounding vehicle trajectory forecasting.
• Integrated TOT prediction and trajectory assessment into the L3-AV decision-making framework.

Main Results:

• The ETSformer model explained over 83% of TOT distribution characteristics.
• Achieved a 0.7% reduction in absolute percentage error for TOT prediction.
• The decision-making framework enabled safe and comfortable optimal decisions.

Conclusions:

• Accurate TOT prediction is critical for L3-AV safety and decision-making stability.
• Understanding TOT's impact improves autonomous driving safety and decision-making techniques.
• The proposed models offer a pathway to more robust and reliable L3-AV systems.